As far as I'm concerned this makes 80% of learning machine learning useless for practical purposes.

With AutoML I don't even have to learn that much about the inner workings of different learning methods. All that I really have to do is create a dataset and specify a target problem, and I can use an AutoML library to figure out what ML architecture I should use to efficiently train models. Even if this won't give me a full workflow it seems like it still removes the majority of the time you'd be spend manually trying to find the right architecture.

All that stuff about how training neural nets is a "black art" that you have to learn by spending months and months building models seems like a waste of time if you're using AutoML.

Am I correct in my assessment?